# Automated blind retrieval evaluation

Run ID: automated-blind-eval-20261028T145147Z
UTC start: 2026-10-28T14:51:47.327890+00:00
UTC finish: 2026-10-28T14:51:48.111334+00:00
Corpus documents: 52
Queries: 21
Independent exact top-5 ordering matches: 42/42
Independently recomputed metric values matching: 168/168
Automated consistency gate: PASS
Human grading required for each run: NO

## Unblinded aggregate results

| Ranker | Precision@5 | Recall@5 | MRR | nDCG@5 | Median query ms |
|---|---:|---:|---:|---:|---:|
| BM25 | 0.3905 | 0.7460 | 0.7897 | 0.6745 | 0.2142 |
| TF-IDF | 0.2762 | 0.5317 | 0.6389 | 0.4337 | 0.1427 |

## Meaning of PASS

PASS means the separately written ranking implementation reproduced every top-five ordering and every scored metric against the fixed, source-based reference fixtures. It creates an automated regression gate without requiring a human to grade every run.

The query/source relevance fixtures are curator-authored and are not independent semantic judgments. Agreement does not prove that the fixtures capture real user intent, that the corpus represents Darkoza/OERSI production, or that a ranker is production-ready. This run changes no production service, index, model weights, or default configuration.

Network access was not requested; no packages were installed. The run uses a local installed system-documentation corpus and standard-library tooling.